AI

Representatives from Nvidia told Illinois representative Bill Foster that it would be technically possible to implement the anti-smuggling measures called for in the Chip Security Act, Foster told MIT Technology Review in an interview on Thursday, even while an industry group claimed they were “potentially infeasible.”  Foster, a Democrat who serves on the Committee on Science, Space, and Technology, is a cosponsor of the act, which would introduce geotracking requirements to verify the location of advanced AI chips in order to prevent their smuggling to China. It is one of three chip-security bills under consideration for inclusion in the Senate’s National Defense Authorization Act, each of which is drawing renewed attention after Nvidia CEO Jensen Huang’s resignation and reported affair with an alleged Chinese spy. When the Chip Security Act was introduced, in May 2025, lawmakers said the technology to implement it existed, something Nvidia has not publicly confirmed. In March 2026, the Semiconductor Industry Association, of which Nvidia is a member, explicitly opposed the bill and wrote that its mandates for chip tracking were “new, untested, and potentially infeasible.”  But privately, Nvidia representatives working with Foster told the lawmaker that the company could implement them.  “They never put anything in writing,” Foster says, but during technical discussions on the possibility of such safeguards, their reaction was “yeah, of course we can do this.” This private acknowledgment to Foster goes further than anything Nvidia has previously made public.  Other technology Nvidia has built for its customers suggests that the company can indeed offer accurate tracking of its chips’ locations. In December 2025, for example, it announced a new tool called Fleet Intelligence that would allow customers to monitor how its advanced chips were performing across different data center locations. In doing so, the tool could indicate which country chips were in, Nvidia confirmed to Reuters, but the company would not elaborate on whether this technology could be used to prevent smuggling.  Nvidia did not respond to requests for comment or confirmation of Foster’s characterization. In early October, when lawmakers were reportedly increasingly concerned about Nvidia’s role in smuggling, MIT Technology Review asked the company if Fleet Intelligence could be used to prevent the practice. It declined to answer directly, instead sending a statement: “Governments and media have found insignificant diversion of controlled US products—a drop in the bucket compared to the ocean of domestic compute China already has.” The recent fallout at Nvidia involves a number of people and companies alleged to be involved in chip smuggling. Alice Huang, with whom Jensen Huang reportedly had an affair, runs Megaspeed, a Singaporean firm now under investigation for smuggling advanced GPUs to China. A US Commerce Department inquiry into the company was reportedly halted after invention by Nvidia. Megaspeed reportedly bought chips from Supermicro, a company that packages Nvidia’s chips onto servers.  The Chip Security Act as drafted would require a mechanism to track the locations of chips, either by inspections of facilities, attestations by companies, or ping-based location mechanisms. The ping-based technique, the hardest option to circumvent, would require the GPU to send a signal of its location when it comes online. In September, the Center for Security and Emerging Technology found it would be the most cost-effective option. Foster said he did not ask Nvidia’s representatives why they weren’t voluntarily adopting such chip-tracking tools if they could technically do so. But if the requirements of the bill pass as part of the Senate’s National Defense Authorization Act, the new tracking mechanisms would only tell investigators where the chips are, not offer a way to disconnect them if it’s revealed they have been smuggled. Foster told MIT Technology Review he intends to introduce new, stricter legislation in the new year that would offer a workaround. The new rule would require licenses to operate advanced GPUs, and before a GPU could connect, its valid license would need to be verified.   In the meantime, the controversy at Nvidia, Foster says, has fueled support among his colleagues for the rules in the bill: “There will be a lot of political pressure, in the wake of everything that’s happened, on the companies to actually do a very good job of the geographic validation of where the chip actually is.”

arXiv:2611.13477v1 Announce Type: cross Abstract: Integrating human feedback into language models has gained increasing attention as a means to align model outputs with human preferences, with Reinforcement Learning from Human Feedback (RLHF) serving as a prominent framework for preference-based alignment. Recent studies have incorporated eye-tracking (ET) data as an additional supervisory signal for reward model training, grounding reward learning in human reading behavior. Building on this approach, we extend prior fixation-centric models by leveraging established findings from cognitive psychology on human letter recognition. Specifically, we model visual attention as a spatially graded distribution centered on the current gaze location and incorporate this representation into reward model training. Our results show that the reward model trained with the proposed gaze distribution achieves higher preference prediction accuracy than a baseline model. Moreover, some results from our experiments are qualitatively consistent with established findings on human attention during reading, suggesting that they potentially capture cognitively meaningful aspects of human attention allocation.

arXiv:2611.13788v1 Announce Type: new Abstract: Minimal-pair benchmarks are only as good as their labels, and LLMs are an obvious substitute for costly human validation. We test it on BLiMP (English) and QFrBLiMP (Quebec French), the two public minimal-pair datasets releasing every forced-choice human judgment. Three LLMs validate by forced choice, by rating, and as a panel, scored against human panels of one, three and five judges in two pre-registered phases. On BLiMP, forced-choice validators agree with held-out human majorities more often than five-judge panels, yet catch far fewer pairs whose label humans reject than one judge does. By point estimate, no validator is worth more than one judge. The errors are not random. On both datasets, across 10 open LMs, human-supported pairs that the pre-registered LLM filter flags are 18.9 points harder than those it accepts, in all 70 LM-validator combinations, and harder than pairs a single human judge flags. Filtering thus removes items on which LMs fail but humans do not, raising accuracy more than random removal. A protocol frozen before any output replicates the gap and the lift on Dutch BLiMP-NL. LLM validation makes minimal-pair benchmarks easier for the models they are meant to test; we release code, validator outputs and item-level scores.

It’s that time of year again: Thousands of delegates, lobbyists, and journalists are gathered for the annual UN climate talks, taking place this time in Antalya, Turkey. COP31 officially kicked off on November 8, marking the beginning of negotiations on one of the biggest stages of the climate world. Turkey and Australia are cohosting this year’s talks, which are scheduled to wrap up on November 20. This annual gathering can often serve as an indicator of where the world stands on addressing and adapting to the climate crisis. And this year’s talks are sure to be a doozy—expect a lot of back and forth on fossil fuels and climate finance in particular. Here’s what you should know about COP31.   In recent years, the tension always feels high when COP rolls around. (COP stands for “Conference of the Parties.”) But this year feels especially charged: The world is still grappling with the ongoing war in Iran, which has disrupted the Middle East and sent economic shock waves around the world as energy prices have spiked. This year’s COP will be the second without a formal delegation from the US. President Trump formally withdrew the country from the Paris Agreement earlier this year and is in the process of withdrawing from the UN Framework Committee on Climate Change. That’s the treaty that allows formal participation in negotiations.  Meanwhile, things aren’t looking good for the Paris Agreement, which was established at UN climate talks 11 years ago, setting a goal of limiting warming to 1.5 °C over preindustrial levels. In a report published in September, the UN Environment Program acknowledged that global warming is set to exceed that 1.5 °C goal, “likely within the next few years.” So as climate talks kick off, there’s an obvious, urgent question looming about how to quickly cut emissions to limit warming. There’s also the major issue of how to pay for efforts to address emissions and to help a warming world endure the inevitable damage.  Two years ago, talks ended in a goal of mobilizing $300 billion per year by 2035 to help pay to reduce emissions and adapt to climate change. (Rich nations provided $136.7 billion in 2024, according to the Organization for Economic Cooperation and Development.) And last year, the COP30 agreement included a goal of tripling the money for adaptation over that period. (Such funding has been quite flat over the past few years.) The world got a tragic reminder of just how high the stakes are earlier this year, when a glacier collapse caused devastating flooding in the Himalayas. The flood killed approximately 1,500 people, left over 5,000 missing, and caused nearly $5 billion in damage. The UN loss and damages fund committed $20 million to helping Nepal recover (less than 1% of the total estimated cost to the country). Another key trend to watch is how the formal negotiations engage with the topic of fossil fuels. Three years ago, the final agreement text explicitly called for a transition away from them. But it felt as if things took a step back last year, when the final agreement didn’t even mention the phrase “fossil fuels” once. This will be an interesting conversation, especially because cohosts Turkey and Australia have both faced criticism over their relationships with these fuels—Australia is a major exporter of coal and natural gas, and Turkey’s grid relies heavily on fossil fuels, including coal. Outside the formal negotiations, COP31 president Murat Kurum, Turkey’s environment minister, is spearheading a side agenda, including two informal initiatives that will be worth watching. These wouldn’t be part of the official agreement text, but countries could sign on to signal support.   One is a pledge to make AI more sustainable; this is almost comically squishy, even as voluntary COP pledges go. The text includes explicit wording that it creates no new obligations for countries, so don’t expect anything resembling enforcement there. The COP leadership is also pushing to make electricity a larger part of member countries’ energy use. Today, only about 23% of global energy is used in the form of electricity. Fossil fuels, like gasoline in vehicles or natural gas for heating, make up the bulk of the rest. Shifting to more electricity use, and relying on cleaner sources for that electricity, could help drive down emissions. The new pledge’s target is for electricity to make up 35% of total energy consumption by 2035.  With leadership billing this as an “implementation COP,” it would be great to see some real, tangible progress on the global challenge of addressing climate change. I’m not expecting anything groundbreaking, but we’ll be tuned in to see what the final agreement shapes up to be.  This article is from The Spark, MIT Technology Review’s weekly climate newsletter. To receive it in your inbox every Wednesday, sign up here. 

arXiv:2611.09402v1 Announce Type: new Abstract: A neural account of lexical access requires an explicit distinction between the structure of lexical representations, the associations linking representational levels, and the activity generated during retrieval. This paper specifies a spiking neural implementation of the mental lexicon within the Neural Engineering Framework and Semantic Pointer Architecture. Conceptual and phonological form representations are constructed recursively from relational graphs, whereas lemma representations provide independent lexical identities. The implementation contains 613 lexical entries and supports two directed processing pathways: perception maps a phonological form to its associated concept via a lemma, whereas production maps a concept to its associated phonological form via a lemma. Each pathway combines two associative-memory variants, winner-take-all (WTA) selection (competition) and thresholded transfer, and uses 199,400 leaky integrate-and-fire neurons. We formalize within-level similarity, cross-level associative retrieval, and the interpretation of decoded activity. Simulations show successful retrieval of eight selected words from the lexicon in both processing directions, with retrieval evaluated against all 613 lexical entries. These simulations use complete phonological forms as input for perception and complete concepts as input for production, illustrating lexical retrieval through the corresponding processing pathways. This paper provides a transparent description of lexical representations and neural retrieval mechanisms, supported by accessible code, lexical data, and online simulation documentation.